Memory Subsystem: Latency

The performance of modern CPUs depends heavily on the cache subsystem. And some applications depend heavily on the DRAM subsystem too. We used LMBench in an effort to try to measure cache and memory latency. The numbers we looked at were "Random load latency stride=16 Bytes".

Mem
Hierarchy
AMD EPYC 7601
DDR4-2400
Intel Skylake-SP
DDR4-2666
Intel Broadwell
Xeon E5-2699v4
DDR4-2400
L1 Cache cycles 4
L2 Cache cycles  12 14-22  12-15
L3 Cache 4-8 MB - cycles 34-47 54-56 38-51
16-32 MB - ns 89-95 ns 25-27 ns
(+/- 55 cycles?)
27-42 ns
(+/- 47 cycles)
Memory 384-512 MB - ns 96-98 ns 89-91 ns 95 ns

Previously, Ian has described the AMD Infinity Fabric that stitches the two CCXes together in one die and interconnects the 4 different "Zeppelin" dies in one MCM. The choice of using two CCXes in a single die is certainly not optimal for Naples. The local "inside the CCX" 8 MB L3-cache is accessed with very little latency. But once the core needs to access another L3-cache chunk – even on the same die – unloaded latency is pretty bad: it's only slightly better than the DRAM access latency. Accessing DRAM is on all modern CPUs a naturally high latency operation: signals have to travel from the memory controller over the memory bus, and the internal memory matrix of DDR4-2666 DRAM is only running at 333 MHz (hence the very high CAS latencies of DDR4). So it is surprising that accessing SRAM over an on-chip fabric requires so many cycles. 

What does this mean to the end user? The 64 MB L3 on the spec sheet does not really exist. In fact even the 16 MB L3 on a single Zeppelin die consists of two 8 MB L3-caches. There is no cache that truly functions as single, unified L3-cache on the MCM; instead there are eight separate 8 MB L3-caches. 

That will work out fine for applications that have a footprint that fits within a single 8 MB L3 slice, like virtual machines (JVM, Hypervisors based ones) and HPC/Big Data applications that work on separate chunks of data in parallel (for example, the "map" phase of "map/reduce"). However this kind of setup will definitely hurt the performance of applications that need "central" access to one big data pool, such as database applications and big data applications in the "Shuffle phase". 

Memory Subsystem: TinyMemBench

To double check our latency measurements and get a deeper understanding of the respective architectures, we also use the open source TinyMemBench benchmark. The source was compiled for x86 with GCC 5.4 and the optimization level was set to "-O3". The measurement is described well by the manual of TinyMemBench:

Average time is measured for random memory accesses in the buffers of different sizes. The larger the buffer, the more significant the relative contributions of TLB, L1/L2 cache misses, and DRAM accesses become. All the numbers represent extra time, which needs to be added to L1 cache latency (4 cycles).

We tested with dual random read, as we wanted to see how the memory system coped with multiple read requests. 

L3-cache sizes have increased steadily over the years. The Xeon E5 v1 had up to 20 MB, v3 came with 45 MB, and v4 "Broadwell EP" further increased this to 55 MB. But the fatter the cache, the higher the latency became. L3 latency doubled from Sandy Bridge-EP to Broadwell-EP.  So it is no wonder that Skylake went for a larger L2-cache and a smaller but faster L3. The L2-cache offers 4 times lower latency at 512 KB. 

AMD's unloaded latency is very competitive under 8 MB, and is a vast improvement over previous AMD server CPUs. Unfortunately, accessing more 8 MB incurs worse latency than a Broadwell core accessing DRAM. Due to the slow L3-cache access, AMD's DRAM access is also the slowest. The importance of unloaded DRAM latency should of course not be exaggerated: in most applications most of the loads are done in the caches. Still, it is bad news for applications with pointer chasing or other latency-sensitive operations. 

Memory Subsystem: Bandwidth Single Threaded Integer Performance: SPEC CPU2006
POST A COMMENT

219 Comments

View All Comments

  • mpbello - Friday, July 14, 2017 - link

    If you are rushing, you reduce scope and deliver fewer pieces with high quality instead of insisting on delivering a full set of benchmarks that you are not sure about its quality.
    The article came to a very strong conclusion: Intel is better for database scenarios. Whatever you do, whether you are rushing or not, you cannot state something like that if the benchmarks supporting your conclusion are not well designed.
    So I agree that the design of the DB benchmark was incredibly weak to sustain such an important conclusion that Intel is the best choice for DB applications.
    Reply
  • StargateSg7 - Sunday, August 06, 2017 - link

    Maybe I'm spoiled, but to me a BIG database is something I usually deal with on a daily basis
    such as 500,000 large and small video files ranging from two megabytes to over a PETABYTE
    (1000 Terabytes) per file running on a Windows and Linux network.

    What sort of read and write speeds do we get between disk, main memory and CPU
    and when doing special FX LIVE on such files which can be 960 x 540 pixel youtube-style
    videos up to full blown 120 fps 8192 x 4320 pixel RAW 64 bits per pixel colour RGBA files
    used for editing and video post-production.

    AND I need for the smaller files, total I/O-transaction rates at around
    OVER 500,000 STREAMS of 1-to-1000 64 kilobyte unique packets
    read and written PER SECOND. Basically 500,000 different users
    simultaneously need up to one thousand 64 kilobyte packets per
    second EACH sent to and read from their devices.

    Obviously Disk speed and network comm speed is an issue here, but on
    a low-level hardware basis, how much can these new Intel and AMD chips
    handle INTERNALLY on such massive data requirements?

    I need EXABYTE-level storage management on a chip! Can EITHER
    Xeon or EPyC do this well? Which One is the winner? ... Based upon
    this report it seems multiple 4-way EPyC processors on waterblocked
    blades could be racked on a 100 gigabit (or faster) fibre backbone
    to do 500,000 simultaneous users at a level MUCH CHEAPER than
    me having to goto IBM or HP for a 30+ million dollar HPC solution!
    Reply
  • PixyMisa - Tuesday, July 11, 2017 - link

    It seems like a well-balanced article to me. Sure the DB performance issue is a corner case, but from a technical point of view its worth knowing.

    I'd love to see a test on a larger database (tens of GB) though.
    Reply
  • philehidiot - Wednesday, July 12, 2017 - link

    It seems to me that some people should set up their own server review websites in order that they might find the unbiased balance that they so crave. They might also find a time dilation device that will allow them to perform the multitude of different workload tests they so desire. I believe this article stated quite clearly the time constraints and the limitations imposed by such constraints. This means that the benchmarks were scheduled down to the minute to get as many in as possible and therefore performing different tests based on the results of the previous benchmarks would have put the entire review dataset in jeopardy.

    It might be nice to consider just how much data has been acquired here, how it might have been done and the degree of interpretation. It might also be worth considering, if you can do a better job, setting up shop on your own and competing as obviously the standard would be so much higher.

    Sigh.
    Reply
  • JohanAnandtech - Thursday, July 13, 2017 - link

    Thank you for being reasonable. :-) Many of the benchmarks (Tinymembench, Stream, SPEC) etc. can be repeated, so people can actually check that we are unbiased. Reply
  • Shankar1962 - Monday, July 17, 2017 - link

    Don't go by the labs idiot
    Understand what real world workloads are.....understand what owning an entire rack means ......you started foul language so you deserve the same respect from me......
    Reply
  • roybotnik - Wednesday, July 12, 2017 - link

    EPYC looks extremely good here aside from the database benchmark, which isn't a useful benchmark anyways. Need to see the DB performance with 100GB+ of memory in use. Reply
  • CarlosYus - Friday, July 14, 2017 - link

    A detailed and unbiased article. I'm awaiting for more tests as testing time passes.
    3.2 Ghz is a moderate Turbo for AMD EPYC, I think AMD could push it further with a higher thermal envelope i/o 14 nm process improvement in the coming months.
    Reply
  • mdw9604 - Tuesday, July 11, 2017 - link

    Nice, comprehensive article. Glad to see AMD is competitive once again in the server CPU space. Reply
  • nathanddrews - Tuesday, July 11, 2017 - link

    "Competitive" seems like an understatement, but yes, AMD is certainly bringing it! Reply

Log in

Don't have an account? Sign up now